I spent years selling programmatic advertising and a meaningful portion of that time in conversations about attribution. Every brand wanted to know what was actually working. Every agency wanted to prove their channel was responsible for the conversion. Every platform had a measurement methodology that showed their channel performing better than the alternatives. The result was an attribution industry built on contested methodologies, platform-reported numbers that couldn't be independently verified, and marketing mix models that could tell you what happened six months ago with enough ambiguity to be interpreted multiple ways.
This wasn't primarily a technology problem. It was a data structure problem, a platform incentive problem, and a fundamental epistemological problem about assigning causality in complex systems. Layering better technology on top of those underlying problems has made attribution more sophisticated without making it more trustworthy. And the current wave of attribution products built on machine learning is, in most cases, repeating the same mistake.
Why attribution is structurally broken, not just technically hard
Attribution's core problem is that the marketing signals we can observe — ad impressions, clicks, email opens, search queries, site visits — are not sufficient to infer causality. A customer who clicks a retargeting ad and then converts was almost certainly going to convert anyway. The click is correlated with conversion, not causal. Last-click attribution assigns 100% of the credit to the retargeting channel. First-touch attribution assigns 100% of the credit to whatever touchpoint happened to be first. Data-driven attribution models distribute credit using statistical weights derived from correlations — which still don't establish causality.
The only way to genuinely measure causal attribution is randomized controlled experiments — exposing some customers to a touchpoint and withholding it from a matched control group, then measuring the difference in outcomes. Marketing mix modeling with media holdouts approximates this. Geo-randomized tests where media runs in some markets and not others approximate it. But these methods require scale, budget for inefficiency (you're deliberately not serving ads to some customers), and careful design. They can't be run continuously across all channels simultaneously. They give you periodic snapshots rather than real-time attribution.
The result is that marketing teams are perpetually making allocation decisions based on attribution data that cannot be trusted as causal. They know this. The best CMOs have always known this. The industry has developed a polite consensus to treat attribution as a proxy metric while managing budget primarily through business outcome proxies — revenue per channel, customer acquisition cost, payback period — that are less precise but less gameable.
What machine learning attribution is actually selling
The AI attribution products on the market today are selling improvements to the correlation-based attribution problem — better statistical modeling, faster computation, more sophisticated multi-touch weighting, integration of more data signals. These are real improvements on the existing methodology. They make correlation-based attribution more nuanced. They are not solving the causality problem.
Some products are going further and incorporating incrementality testing — running automated geo-holdout tests or platform-based lift measurement experiments alongside their attribution models. This is directionally correct. But the incrementality tests are expensive in media efficiency terms (because you're withholding spend in control groups), they require significant statistical power to produce reliable results (which means they need to run for weeks or months, not hours), and they measure incrementality at the campaign or channel level, not at the creative or message level.
We're not saying AI attribution products are useless — they're not. For teams that are currently running on last-click or first-touch models, upgrading to a more sophisticated statistical approach produces real improvement in decision quality. The problem is the marketing: these products are positioned as solving the attribution problem rather than as improving the attribution approximation. That positioning leads to over-reliance on attribution data that still can't tell you what you actually want to know.
The privacy headwind makes the problem worse before it gets better
The trajectory of digital advertising privacy regulation is consistently in the direction of less data available for attribution, not more. Third-party cookie deprecation, mobile identifier restrictions, walled garden measurement restrictions — each of these trends reduces the observable signal available to attribution systems. The models that were calibrated on richer data environments are being asked to run on sparser data, and many are producing less reliable outputs as a result.
The honest answer from most attribution vendors is that their accuracy has degraded over the past three years as privacy changes have taken effect. Some are transparent about this. Others are not. The AI attribution products that will be worth backing are the ones building measurement architectures that are designed for the privacy-constrained environment rather than the open-data environment of 2018 — using first-party data infrastructure, privacy-preserving techniques like differential privacy and aggregated reporting, and incrementality measurement as a first-class methodology rather than an afterthought.
This is a harder product to sell because it requires customers to accept a lower-resolution measurement picture as the honest answer, rather than the high-resolution but unreliable picture that current attribution models provide. But it's the right product for the environment that's coming.
Where AI does meaningfully improve the attribution problem
I want to be honest that AI has genuine applications in measurement and attribution — I'm not arguing the category is worthless. The specific areas where AI adds real value:
Media mix modeling at higher cadence and lower cost. Traditional marketing mix modeling required econometric consultants and months of analysis. AI-assisted MMM can run at faster cycles with less manual effort, making it accessible to mid-size marketing teams that couldn't previously afford it. The methodology is still correlation-based, but the accessibility improvement is real.
Anomaly detection and budget optimization within existing attribution frameworks. AI systems can identify spending patterns that look inefficient relative to historical attribution signals faster than human analysts. This doesn't solve the causality problem but improves the efficiency of the existing allocation process.
Experiment design and power analysis. AI tools that help teams design properly powered incrementality experiments, estimate required sample sizes, and analyze results are genuinely useful. They're accelerating the adoption of the causal measurement methods that are the right long-term answer.
These are real improvements. They matter. The problem is that the market is selling them under the banner of "AI fixes attribution" rather than "AI makes correlation-based attribution better and helps you run more causal experiments." The former is not true. The latter is. The companies we find interesting in this space are the ones leading with the latter.
What we'd want to see from attribution companies we'd back
An honest epistemological posture: clear documentation of what the product can and cannot tell you, and why. The CMOs who stay customers longest are the ones whose expectations were set correctly from the start.
A first-party data architecture strategy: how the product works as the available signal set shrinks. The measurement products built for the 2025 and beyond privacy environment, not the 2019 environment.
Incrementality as a first-class feature, not a bolt-on. The path to causal measurement runs through properly designed experiments, and the product should be accelerating the customer's ability to run them.
Attribution is broken. It's been broken for fifteen years. The companies that have built durable businesses in measurement have been the ones honest about the brokenness while still providing the best available approximation. That's the standard the next generation of attribution products needs to meet.